Solving linear and non-linear SDP by PENNON

Transcription

1 Solving linear and non-linear SDP by PENNON Michal Kočvara School of Mathematics, The University of Birmingham University of Warwick, February 2010

2 Outline Why nonlinear SDP? PENNON the new generation Example: Nearest correlation matrix Example: Approximation by nonnegative splines Example: Sensor network localization

3 Semidefinite programming (SDP) generalized mathematical program min f(x) subject to g(x) 0 A(x) 0 A(x) (non)linear matrix operator R n S m (A(x) = A 0 + x i A i )

4 SDP notations S n... symmetric matrices of order n n A 0... A positive semidefinite A 0... A B 0 A, B := Tr(AB)... inner product on S n A[R n S m ]... linear matrix operator defined by A(y) := n y i A i i=1 with A i S m A [S m R n ]... adjoint operator defined by A (X) := [ A 1, X,..., A n, X ] T and satisfying A (X), y = A(y), X for all y R n

5 Primal-dual SDP pair inf C, X := Tr(CX) (P) X s.t. A (X) = b [ A i, X = b i, i = 1,...,n] X 0 sup b, y := b i y i y,s s.t. A(y) + S = C [ y i A i + S = C] S 0 (D) Weak duality: Feasible X, y, S satisfy C, X b, y = A(y) + S, X y i A i, X = S, X 0 duality gap nonnegative for feasible points

6 Linear Semidefinite Programming Vast area of applications... LP and CQP is SDP eigenvalue optimisation robust programming control theory relaxations of integer optimisation problems approximations to combinatorial optimisation problems structural optimisation chemical engineering machine learning many many others...

7 Why nonlinear SDP? Problems from Structural optimization Control theory Mathematical Programming with Equilibrium Constraints Examples below There are more but the researchers just don t know about...

8 Nonlinear SDP? The general nonlinear SDP (NSDP) problem min f(x) x R n subject to g i (x) 0, h i (x) = 0, A(x) 0 i = 1,..., n g i = 1,..., n h b R n and A : R n S m nonlinear, nonconvex

9

10

11 A380 Inboard Inner Leading Edge Ribs Impact of topological decisions all designs sizing-optimized Baseline Stiffeners at 45-13% -33% Optimized topology

12 Aim: Free Material Optimization Given an amount of material, boundary conditions and external load f, find the material (distribution) so that the body is as stiff as possible under f. The design variables are the material properties at each point of the structure. M. P. Bendsøe, J.M. Guades, R.B. Haber, P. Pedersen and J. E. Taylor: An analytical model to predict optimal material properties in the context of optimal structural design. J. Applied Mechanics, 61 (1994)

13 Free Material Optimization

14 FMO SL primal formulation FMO-problem (minimum volume formulation) min u R n,e 1,...,E m subject to m TrE i i=1 E i 0,ρ TrE i ρ, f u C A(E)u = f i = 1,..., m

15 FMO SL primal formulation FMO-problem with vibration/buckling constraint min u R n,e 1,...,E m subject to m TrE i i=1 E i 0, ρ TrE i ρ, f u C A(E)u = f A(E) + G(E, u) 0 i = 1,..., m

16 FMO SL primal formulation FMO-problem with vibration/buckling constraint min u R n,e 1,...,E m subject to m TrE i i=1 E i 0, ρ TrE i ρ, f u C A(E)u = f A(E) + G(E, u) 0 i = 1,..., m... nonlinear, non-convex semidefinite problem

17 PENNON collection PENNON (PENalty methods for NONlinear optimization) a collection of codes for NLP, (linear) SDP and BMI one algorithm to rule them all READY PENNLP PENSDP PENBMI AMPL, MATLAB, C/Fortran MATLAB/YALMIP, SDPA, C/Fortran MATLAB/YALMIP, C/Fortran NEW PENNON (NLP + SDP) extended AMPL, MATLAB

18 The problem Optimization problems with nonlinear objective subject to nonlinear inequality and equality constraints and semidefinite bound constraints: min f(x, Y) x,y 1 S p 1,...,Y k S p k subject to g i (x, Y) 0, i = 1,...,m g h i (x, Y) = 0, i = 1,...,m h λ i I Y i λ i I, i = 1,...,k. (NLP-SDP)

19 The problem Here x R n is the vector variable Y 1 S p 1,..., Y k S p k are the matrix variables, k symmetric matrices of dimensions p 1 p 1,..., p k p k we denote Y = (Y 1,..., Y k ) f, g i and h i are C 2 functions from R n S p 1... S p k to R λ i and λ i are the lower and upper bounds, respectively, on the eigenvalues of Y i, i = 1,...,k

20 The problem Any nonlinear SDP problem can be furmulated as NLP-SDP, using slack variables and (NLP) equality constraints: write as g(x) 0 g(x) = S S 0 element-wise

21 The algorithm Based on penalty/barrier functions ϕ g : R R and Φ P : S p S p : g i (x) 0 p i ϕ g (g i (x)/p i ) 0, Z 0 Φ P (Z) 0, Z S p. Augmented Lagrangian of (NLP-SDP): i = 1,...,m F(x,Y,u,U,U,p)=f(x,Y)+ m g i=1 u ip i ϕ g(g i (x,y)/p i ) + k i=1 U i,φ P(λ i I Y i ) + k i=1 U i,φ P (Y i λ i I) ; here u R mg and U i, U i are Lagrange multipliers.

22 The algorithm A generalized Augmented Lagrangian algorithm (based on R. Polyak 92, Ben-Tal Zibulevsky 94, Stingl 05): Given x 1, Y 1, u 1, U 1, U 1 ; p 1 i > 0, i = 1,...,m g and P > 0. For k = 1, 2,... repeat till a stopping criterium is reached: (i) Find x k+1 and Y k+1 s.t. x F(x k+1, Y k+1, u k, U k, U k, p k ) K (ii) u k+1 i = u k i ϕ g (g i(x k+1 )/p k i ), i = 1,..., m g (iii) p k+1 i U k+1 = D i A Φ P ((λ i I Y i ); U k i ), i = 1,..., k U k+1 i = D A Φ P ((Y i λ i I); U k i ), i = 1,..., k < p k i, i = 1,..., m g P k+1 < P k.

23 Interfaces How to enter the data the functions and their derivatives? Matlab interface AMPL interface

24 Matlab interface User provides six MATLAB functions: f... evaluates the objective function df... evaluates the gradient of objective function hf... evaluates the Hessian of objective function g... evaluates the constraints dg... evaluates the gradient of constraints hg... evaluates the Hessian of constraints

25 Matlab interface Matrix variables are treated as vectors, using the function svec : S m R (m+1)m/2 defined by a 11 a a 1m a a 2m svec.... sym a mm = (a 11, a 12, a 22,..., a 1m, a 2m, a mm ) T

26 Matlab interface Matrix variables are treated as vectors, using the function svec : S m R (m+1)m/2 defined by a 11 a a 1m a a 2m svec.... sym a mm = (a 11, a 12, a 22,..., a 1m, a 2m, a mm ) T Keep a specific order of variables, to recognize which are matrices and which vectors. Add lower/upper bounds on matrix eigenvalues. Sparse matrices available, sparsity maintained in the user defined functions.

27 AMPL interface AMPL does not support SDP variables and constraints. Use the same trick: Matrix variables are treated as vectors, using the function svec : S m R (m+1)m/2 defined by a 11 a a 1m a a 2m svec.... sym a mm = (a 11, a 12, a 22,..., a 1m, a 2m, a mm ) T Need additional input file specifying the matrix sizes and lower/upper eigenvalue bounds.

28 Example: nearest correlation matrix Find a nearest correlation matrix: min X n (X ij H ij ) 2 (1) i,j=1 subject to X ii = 1, X 0 i = 1,...,n

29 Example: nearest correlation matrix AMPL code: param h{1..21}; set ind within {1..21}; # Nonlinear SDP variables var x{1..21} default 0; minimize Obj: sum{i in 1..21} (x[i]-h[i])ˆ2; subject to l1{i in ind}: x[i] = 1; data; param h:=

30 Example: nearest correlation matrix For H ext = the eigenvalues of the correlation matrix are eigen =

31 Example: nearest correlation matrix The condition number of the nearest correlation matrix must be bounded. Add new variables y, z R and new cosntraints X zi (2) X yi (3) y κz (4) where κ is the required condition number. The constraints do not fit into our required NLP-SDP problem structure. Introduce two new (slack) matrix variables, say, P and Q, and replace (2) and (3) by X zi P = 0 X yi Q = 0 P 0 Q 0

32 Example: nearest correlation matrix More ellegant way: rewrite constraints (2) (3) as I X κi (5) assuming that y = κz and using the transormation of the variable X: z X = X. The new problem: min z, X n (z X ij H ij ) 2 (6) i,j=1 subject to z X ii = 1, I X κi i = 1,...,n

33 Example: nearest correlation matrix For X = the eigenvalues of the correlation matrix are eigen =

34 Example: nearest correlation matrix Cooperation with Allianz SE, Munich: Matrices of size up to Code PENCOR: C code, data in xml format feasibility analysis sensitivity analysis w.r.t. bounds on matrix elements

35 Example: Approximation by nonnegative splines Let f : [0, 1] R. Given its (noisy) function values b i, i = 1,..., n at points t i (0, 1). Find a smooth approximation of f by a cubic spline: 3 P(t) = P (i) (t) = P(i) k (t a i 1 ) k k=1 for a point t [a i 1, a i ], where 0 = a 0 < a 1 <... < a m = 1 are the knots and P (i) k (i = 1,..., m, k = 0, 1, 2, 3) the coefficients of the spline. Spline property: for i = 1,..., m 1 P (i+1) 0 P (i) 0 P(i) 1 (a i a i 1 ) P (i) 2 (a i a i 1 ) 2 P (i) 3 (a i a i 1 ) 3 = 0 (7) P (i+1) 1 P (i) 1 2P(i) 2 (a i a i 1 ) 3P (i) 3 (a i a i 1 ) 2 = 0 (8) 2P (i+1) 2 2P (i) 2 6P(i) 3 (a i a i 1 ) = 0. (9)

36 Example: Approximation by nonnegative splines The function f will be approximated by P in the least square sense: minimize n (P(t j ) b j ) 2 subject to (7),(8),(9). j=1 Now, f is assumed to be nonnegative, so P 0 is required.

37 Example: Approximation by nonnegative splines de Boor and Daniel 74: while approximation of a nonnegative function by nonnegative splines of order k gives errors of order h k, approximation by a subclass of nonnegative splines of order k consisting of all those whose B-spline coefficients are nonnegative may yield only errors of order h 2. Nesterov 2000: P (i) (t) nonnegative there exist two symmetric matrices ( ) ( ) X (i) xi y = i, S y i z (i) si v = i i v i w i such that P (i) 0 = (a i a i 1 )s i (10) P (i) 1 = x i s i + 2(a i a i 1 )v i (11) P (i) 2 = 2y i 2v i + (a i a i 1 )w i (12) P (i) 3 = z i w i (13) X (i) 0, S (i) 0. (14)

38 Example: Approximation by nonnegative splines We want to solve an NLP-SDP problem min P (i) k R j=1 i=1,...,m, k=0,1,2,3 subject to (7),(8),(9), i = 1,..., m (10) (14), i = 1,...,m n (P(t j ) b j ) 2 (15)

39 Example: Approximation by nonnegative splines Example, n = 500, m = 7, noisy data: Figure: Approximation by nonnegative splines: noisy data given in green, optimal nonnegative spline in red and an optimal spline ignoring the nonnegativity constraint in blue. The right-hand side figure zooms on the left valey.

40 Sensor network localization (Euclidean distance matrix completion, Graph realization) We have (in R 2 (or R d )) n points a i, anchors with known location m points x i, sensors with unknown location d ij known Euclidean distance between close points d ij = x i x j, (i, j) I x d kj = a k x j, (k, j) I a Goal: Find the positions of the sensors! Find x R 2 m such that x i x j 2 = d 2 ij, a k x j 2 = d 2 kj, (i, j) I x (k, j) I a

41 Sensor network localization Example, 4 anchors, 36 sensors

42 Sensor network localization Applications Wireless sensor network localization habitat monitoring system in the Great Duck Island detecting volcano eruptions industrial control in semiconductor manufacturing plants structural health monitoring military and civilian surveillance moving object tracking asset location Molecule conformation...

43 Sensor network localization Solve the least-square problem min x i x j 2 d 2 + x 1,...,x m (i,j) I x ij (i,j) I a to global minimum. This is an NP-hard problem. a k x j 2 d 2 kj

44 (P. Biswas and Y. Ye, 04) SDP relaxation Let X = [x 1 x 2... x n ] be a d n unknown matrix. Then x i x j 2 = (e i e j ) T X T X(e i e j ) [ ] a k x j 2 = (a k ; e j ) T Id X T [I d X](a k ; e j ) and the problem becomes (e i e j ) T X T X(e i e j ) = dij 2 ( ) (a k ; e j ) T Id X X T (a Y k ; e j ) = d 2 kj Y = X T X

45 SDP relaxation Now( relax Y = ) X T X to Y X T X, equivalent to Id X Z = X T 0 Y Relaxed problem: min 0 subject to Z 1:d,1:d = I d (0; e i e j ) T Z(0; e i e j ) = d 2 ij (i, j) I x (a k ; e j ) T Z(a k ; e j ) = d 2 kj (k, j) I a Z 0 Full SDP relaxation, FSDP (linear SDP)

46 SDP relaxation Equivalent formulation: min ((0; e i e j ) T Z(0; e i e j ) dij 2 )2 (i,j) I x + ((a k ; e j ) T Z(a k ; e j ) d 2 kj )2 (k,j) I a subject to Z 1:d,1:d = I d Z 0 Full SDP relaxation, FSDP (nonlinear SDP)

47 SDP relaxation Take the SDP solution as initial approximation for the original unconstrained nonconvex problem. Solve both by PENNON.

48 Sensor network localization Example, 9 anchors, 720 sensors

49 Sensor network localization Example, 9 anchors, 720 sensors Figure: SDP: variables, (4 4) LMIs

50 Other Applications, Availability polynomial matrix inequalities (with Didier Henrion) financial mathematics (with Ralf Werner) structural optimization with matrix variables and nonlinear matrix constraints (PLATO-N EU FP6 project) approximation by nonnegative splines approximation of arrival rate function of a non-homogeneous Poisson process (F. Alizadeh, J. Eckstein) sensor network localization (with Houduo Xi) detection of definite pairs of matrices (with F. Tisseur) Many other applications any hint welcome Free academic version of the code available Free downloadable MATLAB version available soon